Twin and family studies show that many common traits and disorders are highly heritable, but genome-wide association studies (GWAS) have been largely unable to identify specific single nucleotide polymorphisms (SNPs) explaining this heritability at the genetic level. Recent work suggests statistical learning methods like gradient boosting (GBM) may be a viable alternative to conventional methods, especially after adjustments for the structure of SNP data. The current research evaluates a two-stage research design for GWAS. GBM is used as a first stage variable selection screen to substantially reduce the dimensionality of SNP data while maintaining sensitivity to additive, nonlinear, and interaction effects, allowing hypothesis testing with a reduced multiple testing burden in the second stage analysis. Thorough simulations shows the proposed two-stage design can substantially improve power to detect effect SNPs in a wide range of conditions. The limitations and potential improvements to this design are explored.